Locally tuned Large Language Model (LLM) to empower digitalization of borehole logs for 3D stratigraphic modelling

Three-dimensional (3D) subsurface stratigraphic modelling is an effective tool to depict site-specific geological conditions for geotechnical design, analysis, and construction. Recent advances in machine learning have enabled 3D stratigraphic modelling in a data-driven manner. However, these methods highly rely on site investigation data (e.g., borehole data), and text descriptions recorded in borehole logs should be interpreted and converted into standardized stratigraphic categories before modelling. In engineering practice, such interpretation and classification tasks are often manually performed by experienced engineers or geologists. This manual workflow is time-consuming, subjective, and sometimes challenging, due to the large volume of borehole data and domain geological knowledge required. This study proposes an innovative paradigm toward automated 3D stratigraphic modelling empowered by a locally deployed large language model (LLM). To integrate domain geological knowledge into existing LLMs, a domain-specific LLM, called stratum generative pre-trained transformer (StratumGPT), is developed and fine-tuned based on a regional dataset of manually classified descriptions from borehole logs. The StratumGPT is then deployed on a local server to ensure data confidentiality and automatically digitalizes text-format stratigraphic descriptions with accuracy comparable to professional judgement by experienced engineers or geologists. The digitalized borehole sticks can be further used as input for data-driven modelling of 3D subsurface stratigraphy. An illustrative case study demonstrated that the proposed LLM-empowered StratumGPT effectively interpreted borehole logs and significantly improved interpretation accuracy, when compared with using a general model directly. The proposed method provides a secure and efficient pathway toward intelligent and automated 3D stratigraphic modelling in engineering practice, with due consideration of data confidentiality.

© 2026 Elsevier Ltd.